The Cloud Above the Clouds: AI and Edge Computing in Space: Prof. Kerri Cahoy

Space Connectivity and Edge Computing Evolution
📌 The future involves cloud above the clouds, moving compute and decision-making on orbit due to established LEO mega-constellations.
🛰️ Current systems rely on ground compute; the shift involves on-orbit decision-making, dynamic tasking, and sending down only necessary insights, reducing data volume.
💡 Enabled by lower launch costs, reusable launch vehicles, and commercial-scale satellite fabrication, coupled with 3GPP standard modifications for necessary latency and power management.

Enabling Technologies and Challenges
🌐 Laser inter-satellite links (Lasercom) are creating a high-capacity space internet spine, comparable to fiber optics on the ground, operating at over 100 Gbps per cross-link.
📡 3GPP releases (5G/6G) are incorporating Non-Terrestrial Networks (NTN) standards to handle user equipment connectivity, compensating for Doppler shifts without requiring changes to mobile phones.
⚠️ Key technical challenges include limited power, thermal constraints (no convection cooling), precision pointing over vast distances, and managing latency and handoffs.

AI, Autonomy, and On-Orbit Operations
🤖 Onboard AI/ML, running on capable commercial processors (GPUs, FPGAs, ASICs), allows for feature recognition, classification, and dynamic tasking based on real-time data analysis.
🔄 Autonomy enables resource scheduling to avoid issues like cloud cover for imaging tasks, optimizing satellite operations based on dynamic, real-time data feeds (e.g., using weather satellite masks).
🩺 Machine Learning models can analyze housekeeping telemetry streams to detect anomalies beyond simple thresholding, inferring issues and potentially enabling self-correction to reduce downtime.

Federated Learning and Future Capabilities
🤝 Federated learning in LEO allows satellite constellations to coordinate and improve a global model by sharing updated weights locally, without transmitting proprietary raw data between vehicles.
🏗️ Autonomy can be used for in-space assembly of larger structures (like bigger antennas) using simplified XYZ Cartesian robots, overcoming volume constraints associated with launching fully assembled systems.
🔮 Future roadmaps target self-healing capability and self-aware connectivity, potentially enabling systems to write software drivers on the fly by interpreting incoming data packets.

Key Points & Insights
➡️ The shift is towards on-orbit inference and decision-making, making space networks behave like terrestrial cloud backbones using Lasercom for high-capacity links.
➡️ Standardization efforts like 3GPP NTN releases are crucial for efficient user device connectivity to satellite platforms, managing complexities like Doppler shifts.
➡️ Federated learning in orbit offers a path for secure, collaborative model improvement across different fleets without compromising sensitive raw data.

📸 Video summarized with SummaryTube.com on Dec 15, 2025, 04:49 UTC